Kronos - A financial candlestick charting platform jointly open-sourced by Microsoft and Tsinghua University

What is Kronos?

Kronos is the first candlestick chart model specifically designed for financial markets, jointly open-sourced by Tsinghua University and Microsoft Research Asia. It predicts future price movements by analyzing candlestick data for assets such as stocks and cryptocurrencies, including opening price, highest price, lowest price, closing price, and trading volume. Kronos employs a two-stage processing framework: an intelligent word segmenter transforms continuous candlestick data into discrete "financial terms," while the large prediction model, based on the Transformer architecture, learns patterns from historical data to predict future trends. The model's training data covers over 45 exchanges globally, enabling it to adapt to the high volatility and noise of financial data.

Kronos offers a variety of parameter models, ranging from 4.1M to 499.2M parameters, to meet diverse needs. It requires only 4 lines of code to load and automatically outputs prediction results after inputting historical data. Kronos provides a real-time BTC/USDT prediction dashboard, integrates Qlib for backtesting validation, and supports A-share market data. On benchmark datasets, Kronos achieves a 93% improvement in RankIC for price series prediction compared to the leading TSFM and an 87% improvement compared to the best non-pre-trained baseline. It achieves a 9% lower MAE in volatility prediction and a 22% improvement in the fidelity of synthetic candlestick sequence generation.

Kronos' main functions

• Candlestick Chart InterpretationKronos can analyze candlestick chart data for financial assets such as stocks and cryptocurrencies, including opening price, highest price, lowest price, closing price, and trading volume, to predict future price trends.
• Two-phase processing frameworkThe system uses an intelligent word segmenter to convert continuous K-line data into discrete "financial terms," and then uses a large-scale prediction model based on the Transformer architecture to learn patterns from historical data for prediction.
• Multiple model optionsIt provides a variety of pre-trained models with parameters ranging from 4.1M to 499.2M to meet different computational and application needs.
• Simple and easy to useThe model can be loaded with only 4 lines of code, and the prediction results will be automatically output after inputting historical K-line data.
• Real-time prediction demonstrationProvides a real-time BTC/USDT prediction dashboard, displaying future trends based on model calculations.
• Support the A-share marketIt integrates Qlib for backtesting and provides a complete fine-tuning pipeline to adapt to your own trading strategies.
• High-performance predictionOn the benchmark dataset, Kronos outperforms the leading TSFM by 93% and the best non-pretrained baseline by 87% in RankIC for price series prediction.

Kronos's technical principles

• Two-phase processing frameworkKronos employs a two-stage processing framework. First, it uses an intelligent word segmenter to transform continuous K-line data into discrete "financial terms." Then, it uses a large-scale prediction model based on the Transformer architecture to learn patterns from historical data and predict future trends.
• Intelligent word segmenterThe word segmenter extracts key information from the K-line data and converts it into discrete symbols that the model can understand and process, providing a foundation for subsequent predictions.
• Transformer architectureThe large-scale prediction model is based on the Transformer architecture, which can process long-series data and capture long-term dependencies in time series, thereby more accurately predicting future price trends.
• Pre-training and fine-tuningKronos offers a variety of pre-trained models that users can fine-tune to better suit specific financial market data and forecasting needs.
• Multi-data source trainingThe model training data covers 45+ exchanges worldwide, enabling it to adapt to the characteristics of financial data in different market environments and possessing strong generalization capabilities.
• Time series modelingKronos focuses on time series modeling and can handle the high volatility and noise of financial data, extracting valuable information for prediction.

Kronos project address

• Github repositoryhttps://github.com/shiyu-coder/Kronos
• arXiv technical paper: https://arxiv.org/pdf/2508.02739

Kronos application scenarios

• Stock market forecastKronos can analyze stock market candlestick charts to predict future stock price movements, helping investors make more informed investment decisions.
• Cryptocurrency tradingApplicable to the cryptocurrency market, it provides price predictions for traders by analyzing cryptocurrency price fluctuations, assisting in the formulation of trading strategies.
• Quantitative trading strategy developmentKronos can serve as an important tool in the development of quantitative trading strategies, providing market trend predictions and optimizing the generation of trading signals.
• Market sentiment analysisBy interpreting candlestick charts, Kronos can reflect changes in market sentiment, helping investors understand the emotional tendencies of market participants and grasp the pulse of the market.
• Risk ManagementBy leveraging its predictive capabilities, investors can better assess investment risks, take proactive risk control measures, and reduce potential losses.
• Financial Data ResearchIt provides financial researchers with a powerful tool to study the patterns of financial markets, price formation mechanisms, and market efficiency.